NVIDIA Kepler speculation thread

[kalelovil]

He was right except about a minor detail: the chip architecture?

He said it would be architecturally part way between Fermi and Kepler, and it is.

 

[silent_guy]

He said it would be architecturally part way between Fermi and Kepler, and it is.

Is it? Does it have different shader architecture? Different texture units? Different ROP? (By know I must have covered the most important characteristics of a GPU.) Different MC? Different crossbar maybe? (You know, that power hog.) Anything else that defines a GPU architecture that comes to mind?

 

[kalelovil]

Is it?

Leaning more towards Fermi than Kepler, sure, but according to Anandtech:

"The codename appears to be GF117 and NVIDIA is keeping many of the details close to their chest, but architecturally it's not simply a die shrink of GF108 and should include some additional enhancements that take advantage of the move to 28nm."

It includes PCI-E 3.0, for example.

Nvidia has kept everything except GK104 under tight wraps this generation. We still don't know the die size of GK107, for instance.

 

[AnarchX]

These are the OEM GDDR5 specs, which are not necessarily the same like the retail specs.

900MHz core and 891MHz (1.8Gbps) DDD3 for GT 640 retail: http://www.4gamer.net/games/160/G016083/20120602002/
(OEM DDR3 was only 793MHz core)

Probably no GDDR5 version?

 

[Richthofen]

Chalnoth:
Yields COULD be bad, because some claims/rumors have it that NV is most favored customer at TSMC. If they get more wafer starts than everybody else, a proportionally greater amount of working chips would be the end result...
*Edit: someone added a bigass post in the time it took me to type two sentences.

Even if Output for NV is higher a shortage could occur. Don't forget that Nvidia needs also more supply accross the board because customers prefer Nvidia over AMD products. Thats the bottom line. Matter of fact is that when the HD7970 launched people were interested to see the benchmarks but not interested to buy one because of 2 reasons:

1) it was quite expensive
2) Nvidia was not there with its counterpart

So most customers were waiting because rumors about the counterpart from Nvidia were flying around. After GK104 launch most of the waiting customers made their choise or decided to wait even further. Looking at the reviews and the value of the Nvidia brand its easy to conclude that demand for GK104 must have been way higher becuase the waiting line that was growing during the 2 - 3 months now actually wanted to buy some product.

 

[KimB]

It's all very relative, I'm not saying GK104 yields are THAT bad, they are simply bad,

What is your reason for thinking this, and what is it relative to?

 

[Sinistar]

What is your reason for thinking this, and what is it relative to?

To quote your fearless leader

The gross margin decline is contributed almost entirely to the yields of 28nm being lower than expected. That is, I guess, unsurprising at this point

Now can you answer the same question you just asked, why you think otherwise?

 

[silent_guy]

To quote your fearless leader

Now can you answer the same question you just asked, why you think otherwise?

That was a quarter before the last one.

To quote him for most recent quarter: "Yields of 28 are probably the best of any new node TSMC has ever done. ... It's yielding wonderfully at this point of the ramp. ... It explains why demand is so great for this particular node."

You're welcome. Any suggestions why I should think otherwise?

 

[Sinistar]

To be honest, I have never seen that quote. Can you provide a link?

 

[ninelven]

How about we just wait for the next earnings report?

 

[A1xLLcqAgt0qc2RyMz0y]

To be honest, I have never seen that quote. Can you provide a link?

http://seekingalpha.com/article/581...3-results-earnings-call-transcript?part=qanda

Click on the Q&A

Jen-Hsun Huang

Yields of 28 are probably the best of any new node that TSMC has ever done. They've done a great job with 28, and I think that it also explains the demand on their 28. It's yielding wonderfully at this point in its ramp. The performance is fabulous. The efficiency, energy efficiency, is terrific. So I think it explains the reason why the demand is so great for this particular node. I think the increased supply will have to come from increased capacity and it's -- we're short now, and every chip that they can come out of their fab is being shipped instantaneously. We've reduced cycle time dramatically between us and our customers, and we're trying to get the products to market as quickly as possible. And so I expect us to be supply constrained from wafers at the wafer level, and we'll continue that way throughout the end of the year.

 

[UniversalTruth]

Another 'woodscrew moment' for dear leader. We can't renounce that his biography is full of such moments.

Where is the Big-K woodscrew puppy?

 

[Sinistar]

So from terrible yields, as noted in their conference call in February, to the best yields ever in May.

I think ninelven is right, I should just wait for the next conference call. They can't legally lie as they could in that article linked.

 

[Alexko]

So from terrible yields, as noted in their conference call in February, to the best yields ever in May.

I think ninelven is right, I should just wait for the next conference call. They can't legally lie as they could in that article linked.

You can't really lie by making subjective statements about "terrible" or "fabulous" yields.

 

[A1xLLcqAgt0qc2RyMz0y]

Another 'woodscrew moment' for dear leader. We can't renounce that his biography is full of such moments.

Where is the Big-K woodscrew puppy?

For someone who has truth in their title you seem unable to handle it when it is presented to you.

28nm yields are great, not enough wafers equals constrained supply.

And what is with all the smiley faces. Is that somehow supposed to make your argument valid?

As for the Big-K, it is not necessary at this point in time as the GK104 currently handles the best AMD has to offer. Why waste wafers on it at this point in time. When wafer supply improves and nVidia can fill all the outstanding demand for current products then the Big-K will be produced.

 

[A1xLLcqAgt0qc2RyMz0y]

So from terrible yields, as noted in their conference call in February, to the best yields ever in May.

I think ninelven is right, I should just wait for the next conference call. They can't legally lie as they could in that article linked.

You seem to be drinking Char-lies Kool-Aid as 'terrible' was never stated nor implied in the previous quarter's CC. The only quote on yields was 'yields of 28-nanometer being lower than expected'. It's only Char-lie who interpreted that as 'terrible'. For all we know Nvidia expected 25% but got 23%.

http://seekingalpha.com/article/370...2-results-earnings-call-transcript?part=qanda

Yields are always lower in the beginning of any new node and always improve over time.

Kevin Cassidy - Stifel, Nicolaus & Co., Inc., Research Division
Maybe along the lines of the 28-nanometer yield. Have you got categorized what the problem is and is there a fix in place? And also, I just -- if you got all the product you needed, do you know what kind of gross margins would have come in?


Jen-Hsun Huang
There's no particular problem. This is the first major quarter of 28-nanometer shipments. There have been some shipments. There've been some shipments in previous quarters but very, very small. And so for TSMC, this is probably the first large quarter of shipments, and we're going to continue to improve yields from here. So there's nothing particularly wrong. This is just early in the learning cycle of a new node. And so we'll improve it with every single outs [ph]. And also this isn't a problem that we can solve. Everybody's using the same 28-nanometer. And so this affects all of us, anybody who uses 28-nanometer. So I think with everybody ramping production, there's going to be a lot more learning cycles both from us and from other people. But TMSC is in a good place now and we just have to keep improving it.

 

[silent_guy]

I think ninelven is right, I should just wait for the next conference call. They can't legally lie as they could in that article linked.

It's the quarterly results conference call, not just an article in a magazine.

 

[Andrew]

So from terrible yields, as noted in their conference call in February, to the best yields ever in May.

You guys are really grasping at straws here to try and find a conspiracy.

The act of knowingly making false statements to investors or the board of directors is a violation of SEC (securities and exchange committee) rules and is a felony under federal law with an average sentence of 5-10 years. That being the case... I think some of you guys should lay off the kool-aid.

I think ninelven is right, I should just wait for the next conference call. They can't legally lie as they could in that article linked.

It's very well known that during Q1 of 2012, the supply of 28nm wafers from TSMC was far less than the demand by companies like Qualcomm, nVidia, and AMD. Qualcomm and nVidia both made public statements regarding how TSMC could not fulfill demand. Perhaps we didn't hear much fuss from AMD because they didn't have nearly as much demand for their cards? My local MIcrocenter and Fry's had 7970's on shelf almost every time I shopped there, and they certainly weren't flying out the door like the gtx 680s have been. Turns out nVidia's sales of the 680 are 60% higher than during the same time period after the GTX 580 was released. And I recall the 580 had big sales.

 

[no-X]

A1xLLcqAgt0qc2RyMz0y: Well, Charlie had interpreted "lower than expected" as "terrible". And you are saying thtat "lower than expected" means:

28nm yields are great

?

Anyway, Nvidia sells 60 % more GTX 680's than GTX 580's. But GTX 580 / GF110 has significantly bigger die, in fact 77 % bigger. That means (if yields are really great), that Nvidia ordered and sold fewer 28nm GK104 wafers than 40nm GF110 wafers. Nvidia expected to launch GK104 as mainstream part, but reserved less wafers than for previous generation high-end part(?) That doesn't make much sense to me.

 

[Sinistar]

The act of knowingly making false statements to investors or the board of directors is a violation of SEC (securities and exchange committee) rules and is a felony under federal law with an average sentence of 5-10 years. That being the case... I think some of you guys should lay off the kool-aid..

I'm pretty sure that is what I implied in my post.